When DDN’s executives sat down with the IT Press Tour in Palo Alto last week, the headline number came early. President and Chief Revenue Officer Kevin Delane told the room that DDN is growing bookings roughly 3x year over year. The company is already running at about 2x through the year, he said, and he expects to finish well beyond 3x. It now counts more than 5,000 customers and has served more than 1 million GPUs.
Growth figures like that get attention on their own. But the more interesting part of the multi-hour briefing was the argument behind them. DDN, with Chief Product Officer Omer Asad doing most of the technical heavy lifting and newly minted Chief Marketing Officer Adam Berlew (with only eight days on the job) jumping in as needed, wants to be known for one thing: making AI infrastructure pay for itself.
“We don’t make GPUs faster. We make them work, profitably.” That was the closing slide, and honestly, it’s the whole talk boiled down in one line.
The Pitch: Demand Is Outrunning Supply
Delane, who has spent more than 30 years in storage and data management and joined DDN six months ago, framed the market bluntly. Demand for AI capacity is outrunning supply, and nobody has enough GPUs, power or floor space. So the job, as he sees it, is to make sure every GPU already bought runs at 99 to 100 percent.
DDN’s presentation leaned on Gartner’s August 2026 numbers to show why the timing matters. Spending on AI-optimized infrastructure as a service hits $42 billion in 2026, nearly double the $21.5 billion baseline of 2025. And 2026 is the first year inference spending ($23.3 billion) passes training ($19 billion). Inference takes 55 percent of the spend this year, with Gartner projecting 59 percent in 2027.
That split shapes everything DDN is building. Training made the company’s name. Inference is where the money is heading.
Same Roots, New Faces
DDN grew up in high-performance computing, and it still carries that DNA. What’s new is the team wrapped around it. Asad joined just over a year ago from AWS. Delane came over from Salesforce. The CFO, Guido Torrini, has been there about nine months, and Rick Scurfield now runs growth and the partner ecosystem as Chief Growth Officer. Berlew arrived barely a week before the briefing, fresh from Equinix in the Netherlands.
Why does that matter? Because DDN has a branding gap, and the executives said so themselves. Plenty of Google Cloud customers have no idea DDN is running underneath their training jobs.
Berlew’s marching orders are clear: start asking customers for permission to use their logos, then start publicizing. “Next time we meet, you’ll see more logos,” he promised.
The customer base is already broad, though. Here is how DDN describes it:
- NVIDIA Cloud Partners and AI clouds, where large transactions keep growing.
- AI-native and sovereign AI builders, including Shakti Cloud in India.
- Traditional HPC strongholds like energy, aerospace, financial services, government, higher ed and healthcare.
- Enterprise AI, which Delane described as a “land and expand” play.
Two Partners Doing the Heavy Lifting
If you want to understand DDN, start with NVIDIA and Google.
NVIDIA has run its own AI clusters on DDN since 2016. More than 9,000 GPUs sit on DDN storage, including the EOS and Selene SuperPODs (12 PB and 15 PB). Asad said Nemotron, NVIDIA’s open-weights model family, is trained on DDN, and that the reference architecture NVIDIA leases at neo clouds is DDN too. The slide deck carried a Jensen Huang quote: “Without DDN, NVIDIA supercomputers wouldn’t be possible.” That quote alone is worth 40 customer logos.
Google is the other half. EXAScaler powers Google Cloud Managed Lustre, and Asad was quick to say this is not a marketplace listing. It’s a first-party service. A customer clicks “managed training” in the Google Cloud console, DDN fires up underneath, Google collects the money, and DDN gets paid on the back end. Google’s go-to-market machine does the selling, which Delane called “light touch” for DDN, and customers get added daily. DDN’s numbers for that service: 1.5x faster model training and 42 percent lower training costs.
“Before You Hit Max Power, You’re Already Bottlenecked at Data”
This was Asad’s central argument, and it was a good one. Every neo cloud and frontier lab starts with a power envelope: 50 megawatts, 500 megawatts, a gigawatt. They pay for it, then roll in the GPU racks. But Asad’s point was that the data layer chokes the GPUs before they ever reach that power ceiling.
“You bought a 747. Keep it flying. Don’t keep it parked.”
He walked through how different each stage of the AI pipeline looks to a storage system:
Training Wants Parallel Reads, Then Suddenly Wants Writes
Training hammers storage with huge parallel reads. Then GPUs fail (more often than vendors like to admit, Asad noted), so jobs stop to write checkpoints. That flips the pattern to bursty, erratic writes, and thousands of GPUs sit idle until the last one gets its acknowledgment from storage.
Inference Is a Different Animal
Now you get random reads of unpredictable sizes, models loaded and reloaded on demand, and many tenants sharing the same hardware. Add data preparation on top, where you might need to tag and list billions of objects in a single 50 PB bucket, and you see the problem. Different stages have competing demands, and running a separate vendor for each one multiplies cost and administrators.
His summary: “At eight nodes, everybody shines.” The hard part is staying fast and simple at hundreds of petabytes.
EXAScaler: The Training Workhorse
DDN’s answer for training is EXAScaler, its parallel file system built on Lustre but, as Asad stressed, far from plain Lustre. DDN bought Whamcloud and says it is among the biggest contributors to the open-source project; Asad put the share of Lustre developers on DDN’s payroll at 70 to 80 percent. But open-source Lustre lacks multi-tenancy and quality of service, he said, and takes a pile of engineering effort to run. EXAScaler is the push-button version.
The architecture is the point. Every compute client talks to every storage server in parallel, with no gateway in the path. DDN’s slides put numbers on what that buys:
- 190 GB/s and 8M IOPS from one 2U building block drawing 2.5 kW.
- Up to 99 percent GPU busy time and 96 percent network efficiency in NVIDIA’s DGX SuperPOD B200 reference design.
- 15x faster checkpointing, so GPU hours go to training instead of saving state.
- EOS, NVIDIA’s own supercomputer: 576 DGX systems on fewer than three racks of DDN storage, using about 100 kW for 18.4 exaflops.
Asad also claimed that in head-to-head tests on 50,000-GPU clusters, DDN finishes jobs in half the time about 80 percent of the time, and called EXAScaler eight times faster than its nearest competitor. Those are vendor claims, of course. His response to skepticism? “Any time, anywhere, test that.”
He also pointed to the scale of frontier-lab deployments. Mistral, Nemotron and Grok were all trained on DDN, he said, with a single Mistral environment hitting DDN from 100,000 GPUs through a 4U controller. Shakti Cloud runs 8,000 NVIDIA B200 GPUs on EXAScaler and reports 99 percent GPU busy time across tenants.
Infinia: The Four-Year Bet on Inference
Here’s where the story gets newer, and where I think readers should pay closest attention. Infinia took DDN four years to build, and it is not a file system. It’s a distributed key-value data layer, closer in spirit to the data systems behind Amazon’s Dynamo or Google’s Spanner, with metadata and data scaling together as you add nodes. You reach it through S3, POSIX files, a native KV cache interface or a public SDK, and all of them sit on the same SDK underneath.
Why Context Is the Money Problem
When a model answers a question, it builds a KV cache, a running memory of the context. GPU memory fills up, and if that context has to be recomputed, you pay for GPU time twice. DDN puts a number on it: on a 4,096-GPU B200 fleet serving real chat traffic at median on-demand pricing, waiting costs $186,000 in lost revenue every day. Asad noted the model can be reproduced by replaying public Mooncake inference traces against your own system.
Infinia keeps that cache outside the GPU and hands it back over RDMA, straight to GPU memory. The integration ships inside NVIDIA’s NIXL and the Dynamo container, and DDN says one integration covers Dynamo, vLLM, SGLang and LMCache. Asad said DDN was among the first vendors to open source and ship inside Dynamo. The result: 4.2x faster time to first token at 100K-token context.
Does It Hold Up at Scale?
DDN offered several data points:
- 300-plus nodes in a single namespace at a frontier lab, serving about 2.3 TB/s of GET/PUT traffic.
- A 700 PB production deployment across four data halls, with peak GET throughput of 6.4 TB/s and 5M small-object GETs per second. DDN operates it as a managed service for a frontier lab, and Asad said it serves roughly 200,000 GPUs from one namespace, without KV cache optimization turned on.
- Audited object performance on Oracle Cloud Infrastructure of 225K GET and 194K LIST requests per second.
Multi-tenancy got equal billing. One API call sets a tenant’s keys, quota and access, and weighted fair share in the data path keeps one tenant’s spike from hurting the others. Asad’s analogy was a good one: “If there is no car on the road, go nuts. But if there are other cars, stay in your lane.”
A fair note here. Infinia is young. Asad said about 15 deployments are in production, ranging from 7 nodes to 300-plus.
One Fabric, Two Engines
So does Infinia eventually replace EXAScaler? A tour colleague asked twice, and the answer was no both times. EXAScaler stays the engine for training at 100,000-GPU scale. Infinia leads when object is the front door, which Asad said covers about 99 percent of data prep and inference use cases. The two exchange data, and a customer on Infinia who needs a file interface can bring in EXAScaler within the same namespace.
Tying them together is DDN’s Horizon layer (Control, Store, Manage and Support), which is supposed to give operators one set of interfaces and APIs across both engines. Everything is software-first, running on qualified hardware from Supermicro, Dell, Lenovo and others, or DDN appliances.
Roadmap items worth noting:
- Native Kafka and vector database interfaces next year.
- Token-priority hints, co-developed with NVIDIA, so heavy or important contexts aren’t thrown away under contention.
- Continued work with AMD and Google TPUs. Asad said DDN “jumps in right away” wherever optimization libraries exist, though Delane was candid that NVIDIA is the fastest path to growth.
DDN Runs on Its Own AI, Too
I liked this part. DDN streams telemetry from customer systems into its data lakes and runs locally trained models over it to spot training stalls, GPU failures and hot spots faster than a human could. Asad said the models are fine-tuned open models (he named Kimi), with Claude licenses used for retraining and analysis. Developers use AI for code generation, review and test creation. DDN is now on a release cycle shorter than three months.
The Part Nobody Can Ignore: Power
The conversation wandered, as the best ones do, into data center power, water, permits and community pushback. Berlew, newly out of Equinix, said every provider is hunting for alternatives, but the shortage is real and raising the output of GPUs already deployed matters now. Delane added that building outside the US is easier, which is why Asia Pacific is growing faster for DDN than the US or Europe. Reporters around the table argued about token rationing, grid limits in the Netherlands and whether sustainability still sells. Nobody had a tidy answer.
That’s precisely why DDN’s argument lands. If you can’t get more power, you’d better get more work out of the power you have. Asad framed it this way: serving tokens faster is great, but “at what rack capacity? At what power capacity?”
Why You Should Pay Attention
Delane was asked about outside investment plans and stuck to the growth story. Fair enough; the numbers he gave were strong. But DDN’s real case isn’t about funding. It rests on three things that are hard to fake: a decade inside NVIDIA’s own clusters, a first-party spot in Google Cloud, and production deployments in the hundreds of petabytes.
Is every performance claim something you should take at face value? No. Run your own tests, as Asad invited. Still, if your GPUs are the most expensive line item in your data center and they spend part of every day waiting on data, DDN has an answer worth hearing out. After this briefing, I’d put it near the top of the list of infrastructure vendors to watch as inference takes over the AI budget.
##
